base/asm-common.h (x86), and knock-on: Add macros for full-size regs.

[catacomb] / math / mpx.c

diff --git a/math/mpx.c b/math/mpx.c
index 18baf2f..07a6c20 100644 (file)
--- a/math/mpx.c
+++ b/math/mpx.c
@@ -545,15 +545,21 @@ MPX_SHIFTOP(lsr, {
   size_t nr = MPW_BITS - nb;
   mpw w;
 
-  av += nw;
-  w = av < avl ? *av++ : 0;
-  while (av < avl) {
-    mpw t;
-    if (dv >= dvl) goto done;
-    t = *av++;
-    *dv++ = MPW((w >> nb) | (t << nr));
-    w = t;
+  if (nw >= avl - av)
+    w = 0;
+  else {
+    av += nw;
+    w = *av++;
+
+    while (av < avl) {
+      mpw t;
+      if (dv >= dvl) goto done;
+      t = *av++;
+      *dv++ = MPW((w >> nb) | (t << nr));
+      w = t;
+    }
   }
+
   if (dv < dvl) {
     *dv++ = MPW(w >> nb);
     MPX_ZERO(dv, dvl);
@@ -812,7 +818,7 @@ void mpx_usub(mpw *dv, mpw *dvl, const mpw *av, const mpw *avl,
 
 void mpx_usubn(mpw *dv, mpw *dvl, mpw n) { MPX_USUBN(dv, dvl, n); }
 
-/* --- @mpx_uaddnlsl@ --- *
+/* --- @mpx_usubnlsl@ --- *
  *
  * Arguments:  @mpw *dv, *dvl@ = destination and first argument vector
  *             @mpw a@ = second argument
@@ -917,19 +923,25 @@ static void simple_umul(mpw *dv, mpw *dvl, const mpw *av, const mpw *avl,
 
 #if CPUFAM_X86
   MAYBE_UMUL4(x86_sse2)
+  MAYBE_UMUL4(x86_avx)
 #endif
 
 #if CPUFAM_AMD64
   MAYBE_UMUL4(amd64_sse2)
+  MAYBE_UMUL4(amd64_avx)
 #endif
 
 static mpx_umul__functype *pick_umul(void)
 {
 #if CPUFAM_X86
+  DISPATCH_PICK_COND(mpx_umul, maybe_umul4_x86_avx,
+                    cpu_feature_p(CPUFEAT_X86_AVX));
   DISPATCH_PICK_COND(mpx_umul, maybe_umul4_x86_sse2,
                     cpu_feature_p(CPUFEAT_X86_SSE2));
 #endif
 #if CPUFAM_AMD64
+  DISPATCH_PICK_COND(mpx_umul, maybe_umul4_amd64_avx,
+                    cpu_feature_p(CPUFEAT_X86_AVX));
   DISPATCH_PICK_COND(mpx_umul, maybe_umul4_amd64_sse2,
                     cpu_feature_p(CPUFEAT_X86_SSE2));
 #endif
@@ -1262,6 +1274,7 @@ mpw mpx_udivn(mpw *qv, mpw *qvl, const mpw *rv, const mpw *rvl, mpw d)
 
 #include <mLib/alloc.h>
 #include <mLib/dstr.h>
+#include <mLib/macros.h>
 #include <mLib/quis.h>
 #include <mLib/testrig.h>
 
@@ -1354,7 +1367,7 @@ static int loadstore(dstr *v)
       ok = 0;
     MPX_OCTETS(oct, m, ml);
     mpx_storel(m, ml, d.buf, d.sz);
-    if (memcmp(d.buf, v->buf, oct) != 0) {
+    if (MEMCMP(d.buf, !=, v->buf, oct)) {
       dumpbits("\n*** storel failed", d.buf, d.sz);
       ok = 0;
     }
@@ -1364,7 +1377,7 @@ static int loadstore(dstr *v)
       ok = 0;
     MPX_OCTETS(oct, m, ml);
     mpx_storeb(m, ml, d.buf, d.sz);
-    if (memcmp(d.buf + d.sz - oct, v->buf + v->len - oct, oct) != 0) {
+    if (MEMCMP(d.buf + d.sz - oct, !=, v->buf + v->len - oct, oct)) {
       dumpbits("\n*** storeb failed", d.buf, d.sz);
       ok = 0;
     }
@@ -1398,14 +1411,14 @@ static int twocl(dstr *v)
 
     mpx_loadl(m, ml0, v[0].buf, v[0].len);
     mpx_storel2cn(m, ml0, d.buf, v[1].len);
-    if (memcmp(d.buf, v[1].buf, v[1].len)) {
+    if (MEMCMP(d.buf, !=, v[1].buf, v[1].len)) {
       dumpbits("\n*** storel2cn failed", d.buf, v[1].len);
       ok = 0;
     }
 
     mpx_loadl2cn(m, ml1, v[1].buf, v[1].len);
     mpx_storel(m, ml1, d.buf, v[0].len);
-    if (memcmp(d.buf, v[0].buf, v[0].len)) {
+    if (MEMCMP(d.buf, !=, v[0].buf, v[0].len)) {
       dumpbits("\n*** loadl2cn failed", d.buf, v[0].len);
       ok = 0;
     }
@@ -1442,14 +1455,14 @@ static int twocb(dstr *v)
 
     mpx_loadb(m, ml0, v[0].buf, v[0].len);
     mpx_storeb2cn(m, ml0, d.buf, v[1].len);
-    if (memcmp(d.buf, v[1].buf, v[1].len)) {
+    if (MEMCMP(d.buf, !=, v[1].buf, v[1].len)) {
       dumpbits("\n*** storeb2cn failed", d.buf, v[1].len);
       ok = 0;
     }
 
     mpx_loadb2cn(m, ml1, v[1].buf, v[1].len);
     mpx_storeb(m, ml1, d.buf, v[0].len);
-    if (memcmp(d.buf, v[0].buf, v[0].len)) {
+    if (MEMCMP(d.buf, !=, v[0].buf, v[0].len)) {
       dumpbits("\n*** loadb2cn failed", d.buf, v[0].len);
       ok = 0;
     }